WeeLLM dynamically streams text encoder and transformer layers to the GPU for massive models. Instead of forcing the entire model into VRAM, it intelligently pins as many blocks as your hardware allows, and seamlessly streams the rest layer-by-layer in the background — enabling massive models to run smoothly on budgets as low as 4GB VRAM.
🚀 Official ComfyUI Extension Now Available!
Bring WeeLLM's ultra-low VRAM layer-streaming capabilities directly into your ComfyUI workflows!
| Model | Parameters | Peak VRAM | Peak RAM | Time (1024x1024) |
|---|---|---|---|---|
SDXL (Juggernaut XL v9) |
~6.6B | 2.98 GB | 1.50 GB | ~120s · 20 steps |
SD 1.5 |
~1.7B | 0.90 GB | 1.40 GB | ~68s · 20 steps |
SD 3.5 Medium |
~8B | 3.48 GB | 3.96 GB | ~219s · 20 steps |
FLUX.1-dev |
~12B | 1.51 GB | ~2.0 GB | — |
FLUX.1-Kontext-dev |
~12B | 1.51 GB | ~2.0 GB | — |
FLUX.1-schnell |
~12B | 1.64 GB | 1.68 GB | ~159s · 4 steps |
FLUX.2-klein-4B |
4B | 2.0 GB | 2.3 GB | ~61s · 4 steps |
Lumina-Image-2.0 |
~2B | 1.38 GB | — | — |
CogView4-6B |
~15B | 2.44 GB | 1.92 GB | ~411s · 10 steps |
Z-Image-Turbo |
~10B | 1.60 GB | 1.70 GB | ~167s · 4 steps |
HiDream-I1-Full |
~15B | 3.68 GB | 2.02 GB | ~797s · 10 steps |
LongCat-Image |
~9B | 2.19 GB | — | ~200s · 10 steps |
Qwen-Image (Qwen/Qwen-Image) |
~20B | 2.47 GB | 1.60 GB | ~795s · 10 steps |
AuraFlow (fal/AuraFlow) |
~4B | 1.34 GB | 1.61 GB | ~405s · 10 steps |
ERNIE-Image (Baidu/ERNIE-Image) |
~10B | 1.69 GB | 2.54 GB | ~123s · 5 steps |
Krea-2-Turbo |
~13B | 3 GB | 3.23 GB | ~810s · 10 steps |
MiniMax-H3 |
~34B | 3.14 GB | — | — |
MOSS-Image-SFT |
~14B | 1.25 GB | 2.10 GB | ~260s · 6 steps |
The models run with no quantization on RTX-3050 — full bfloat16 (1024x1024) weights streamed layer-by-layer to GPU.
- SDXL
- SD 1.5
- SD 2.0
- SD 2.1
- SD 3 Medium
- SD 3.5 Medium
- SD 3.5 Large
- Flux.1 Dev
- Flux.1 Schnell
- Flux.1 Kontext
- Flux.1 Krea
- Flux.2 Klein 4B
- Flux.2 Klein 9B
- Flux.3 Klein (Base + Distilled)
- CogView 4
- LongCat Image